music_test.X = z_norm_by_feature(music_test.X, mean_X, std_X)
# Balacing train data.
# print "Balacing train data."
# music_train.balance_data_oversampling_smote_regular()
# Set train parameters.
# lambdav = 0.00001
lambdav = 0
# alpha = 0.0000001
# iterations = 1000000
alpha = 0.1
iterations = 1200
# print "Solving normal equation."
theta = solve_normal_equation(music_train.X, music_train.y, lambdav)
print "Solving using gradient descent."
# theta = gradient_descent(music_train.X, music_train.y, None, alpha, lambdav, iterations)
#theta, J_history = gradient_descent_with_J_history(music_train.X, music_train.y, None, alpha, lambdav, iterations)
#plot_history(J_history)
print "Computing cost."
print compute_cost(music_train.X, music_train.y, theta, lambdav)
print compute_cost(music_validation.X, music_validation.y, theta, lambdav)
print compute_cost(music_test.X, music_test.y, theta, lambdav)
for delta_year in range(10):
print delta_year
print "Computing train accuracy."
for year2 in range(2000, 2010):
print year1, year2
delta_year = 5
less_year = music_train.y <= year1
less_year.shape = (len(music_train.y))
greater_year = music_train.y > year1
greater_year.shape = (len(music_train.y))
music_train_y_year_yes_or_not = np.array(music_train.y)
music_train_y_year_yes_or_not[less_year] = year1
music_train_y_year_yes_or_not[greater_year] = year2
music_train_y_year_yes_or_not.shape = (
len(music_train_y_year_yes_or_not), 1)
# < year or > year classifier.
theta_year_yes_or_not = solve_normal_equation(
music_train.X, music_train_y_year_yes_or_not, 0)
# < year classifier.
y = np.array(music_train.y[less_year])
y.shape = (len(y), 1)
X = music_train.X[np.where(less_year)]
theta_year_less = solve_normal_equation(X, y, 0)
print compute_accuracy(X, y, theta_year_less, delta_year)
# > year classifier.
y = np.array(music_train.y[greater_year])
y.shape = (len(y), 1)
X = music_train.X[np.where(greater_year)]
theta_year_more = solve_normal_equation(
music_train.X[greater_year], y, 0)
print compute_accuracy(X, y, theta_year_more, delta_year)
#music_train.balance_data_undersampling_cluster_centroids()
#music_train.balance_data_undersampling_tomek_links()
music_train.balance_data_ensemblesampling_balance_cascade()
#music_train.balance_data_ensemblesampling_balance_cascade()
after_balacing_size = len(music_train.X)
print "Before balacing size: " + str(before_balacing_size)
print "After balacing size: " + str(after_balacing_size)
# Set train parameters.
lambdav = 0.0000000001
n = len(music_train.X[0])
print "Solving normal equation."
# Get thetas to reduce data.
theta = solve_normal_equation(music_train.X, music_train.y, lambdav)
ordered_theta = np.argsort(np.abs(theta).reshape(len(theta)))
ordered_theta = ordered_theta[::-1]
# Initialize costs.
J_history_train = np.zeros(n)
J_history_validation = np.zeros(n)
for iteration in range(n):
theta = solve_normal_equation(music_train.X[:, ordered_theta[:(n - iteration)]], music_train.y, lambdav)
J_history_train[iteration] = compute_cost(music_train.X[:, ordered_theta[:(n - iteration)]], music_train.y, theta, 0)
J_history_validation[iteration] = compute_cost(music_validation.X[:, ordered_theta[:(n - iteration)]], music_validation.y, theta, 0)
print "Theta size: " + str(n - iteration)
print "J_train: %f" % J_history_train[iteration]
print "J_validation: %f" % J_history_validation[iteration]
